The real switching cost is memory

Every time you switch model providers using their own AI tools, you leave behind the knowledge you built inside them. That’s the cost nobody is pricing in.

In the past few months alone, 1.5 million people pledged to leave ChatGPT over values misalignment. Anthropic lost its government contracts, then cut third-party tool access for subscribers. A class action lawsuit alleges that Perplexity secretly shared user conversations with Meta and Google, including prompts typed in incognito mode. New models and capabilities launch weekly, pulling users toward different platforms. Switching AI tools is no longer unusual. It’s becoming a regular part of how people and organisations use this technology.

And we’re still early. The capabilities available today are a fraction of what’s coming, and nobody can predict which provider will lead in six months. Being locked into a single tool when the landscape is this fluid isn’t just inconvenient. It’s a strategic constraint at a time when flexibility matters most.

Interested in learning more about what’s actually possible with AI and the options for family offices?

Why people switch

Three drivers keep surfacing. Capability: a new model does something yours can’t. Deeper reasoning, computer use, better code generation, longer context. Values: the provider’s relationship with governments, their stance on surveillance, or their corporate decisions no longer align with yours. Trust: terms change quietly, data handling isn’t what was promised, or a lawsuit surfaces that reveals practices the user never consented to.

For family offices, all three are relevant. But trust carries particular weight. These are organisations built on discretion, confidentiality, and long-term relationships. When a tool breaks that trust, the switch isn’t gradual. It’s immediate. There’s no phased transition when you discover your data was shared with ad platforms. There’s no “let’s evaluate alternatives over the next quarter” when a provider’s government entanglements create exposure you didn’t sign up for. The decision is made fast, and the move happens overnight.

What you lose when you leave

The subscription transfers easily. The memory doesn’t. Six months of conversation history with an AI tool contains portfolio reasoning, compliance logic, communication preferences, relationship context, decision patterns. That accumulated knowledge doesn’t export. It stays with the provider. You start the next tool from zero.

The more context you build inside a tool, the harder it becomes to leave, not because of price, but because of what you’d lose.

The onboarding work, the context-building, the months of accumulated understanding: all of it begins again. For a family office that has been building institutional memory inside an AI tool, this isn’t a minor inconvenience. It’s the loss of something the office had never managed to capture before, lost precisely because it was captured in the wrong place.

The practical options and their limits

There are workarounds. You can ask the current model to summarise your key context, preferences, and patterns into a document you can feed into the new tool. Some people export conversation histories. Others manually write onboarding prompts for the replacement. These help, but they lose fidelity. The nuance, the accumulated judgment, the patterns the model learned through months of interaction don’t transfer cleanly through a summary. You get a compressed version of what you had, not the thing itself.

For offices that value trust highly and know that a breach of that trust could force a rapid switch, this gap is worth planning for now. Having a backup plan that makes switching as seamless as possible isn’t paranoia. It’s the same operational readiness the office applies to every other critical dependency.

There's no phased transition when you discover your data was shared with ad platforms.

The architecture that solves this

The deeper answer is infrastructure. If the office’s memory layer sits independently of any single provider, connecting to whichever model the office chooses to use, the switching cost drops to near zero. The context is portable. The model becomes interchangeable. This is the same argument that runs through the sovereign AI conversation: own your infrastructure, own your data, own your memory.

Third-party tools and middleware platforms are beginning to build for exactly this, allowing organisations to maintain a persistent knowledge layer that connects to different models underneath. The offices that design their AI architecture this way will be able to switch providers without losing what they’ve built. The ones that don’t will find themselves staying with a provider longer than they should, not because the tool is still the best option, but because leaving costs too much.

Memory as the new lock-in

The subscription era trained us to think switching costs were financial. In AI, they’re cognitive. The more context you build inside a tool, the harder it becomes to leave, not because of price, but because of what you’d lose. For family offices, where the knowledge is sensitive, personal, and hard to reconstruct, that cost is higher than for most.

The offices that recognise this now, that treat memory architecture as a design decision rather than an afterthought, will have the freedom to move when they need to. And in a landscape where trust can break overnight, capability shifts weekly, and the tools you rely on today may not be the tools you want tomorrow, that freedom is worth designing for.